Scholarlink

<p>The Antarctic ozone hole constitutes a critical global environmental issue. Although signs of recovery have emerged since the early 21st century, ongoing climate change has heightened uncertainty about its future trajectory. This study uses ERA5 as a reference dataset and ozone simulations from 13 models participating in the Coupled Model Intercomparison Project Phase 6 (CMIP6), aiming to improve Antarctic ozone projections through the optimized model screening and Multi-Model Ensemble (MME) approach. Three novel SHAP (SHapley Additive exPlanations) value screening methods derived from Tree-Based models, along with one traditional screening method, are developed for CMIP6 ensemble selection. Based on these methods, five machine learning (ML) models are further constructed for future Antarctic ozone projections. The main conclusions are as follows: (1) All CMIP6 models show general consistency with ERA5 but exhibit varying degrees of bias. Even the best-performing model, CESM2, performs well in terms of Taylor Skill Score (TSS) and Root Mean Square Error (RMSE), but shows notable deficiencies in Interannual Variability Skill Score (IVS); (2) Among the four screening methods, the SHAP-based Worst-to-Best Screening, which progressively removes poorly performing models, achieves the best performance. The optimal Artificial Neural Network (ANN) model based on this method (MME6-ANN) significantly improves the MME (<i>R</i> = 0.833, IVS × 10 = 0.001, Mean Squared Error (MSE) = 200.81), substantially outperforming the traditional equal-weighted ensemble (MMM13) derived from 13 CMIP6 models and achieving the best agreement with the reference; (3) The Antarctic ozone recovery years projected by MME6-ANN under the four Shared Socioeconomic Pathway (SSP) scenarios (SSP1-2.6, SSP2-4.5, SSP3-7.0, and SSP5-8.5) are 2065, 2062, 2057, and 2056, respectively, representing delays of 9, 11, 7, and 7 years compared with the traditional MME. Overall, this study offers a more reliable estimate of future Antarctic ozone recovery under climate change, and provides a robust reference for CMIP6 MME projections using machine learning.</p> Graphical Abstract <p></p> <p>The graphical abstract presents a machine learning–based enhanced CMIP6 model screening and multimoded ensemble framework (MME-ML) for improving Antarctic ozone future projections. ERA5 reanalysis data and ozone simulations from CMIP6 models were first processed into monthly time series, which are then divided into training, validation, and test sets. Subsequently, three innovative screening methods based on SHAP values derived from Tree-Based models, alongside one traditional screening method based on evaluation metrics including the TSS, IVS, and RMSE, were applied to screen the CMIP6 models. Based on the screened CMIP6 model ensembles from the four screening methods, five machine learning models (GLM, RF, XGBoost, LSTM, and ANN) were constructed, and a comprehensive evaluation identified the optimal screening method and its corresponding optimal ML model. Finally, the optimal ML model was used to improve the MME, correcting the CMIP6 Antarctic ozone projections under four SSP scenarios. The results indicate that the Antarctic ozone recovery years projected by the improved MME under the four scenarios (SSP1-2.6, SSP2-4.5, SSP3-7.0, and SSP5-8.5) are 2065, 2062, 2057, and 2056, respectively, representing delays of 9, 11, 7, and 7 years compared with the traditional MME. This research yields more accurate estimates of future Antarctic ozone recovery under climate change, and provides a transferable framework for CMIP6 multimodel ensemble projections using machine learning techniques.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Machine Learning-based Optimization of CMIP6 Multimodel Ensemble for Antarctic Stratospheric Ozone Projections

  • 期刊论文

  • 发表日期：

    2026年5月18日

查看全文
  • Junzhe Chen,
  • Yu Zhang,
  • Houxiang Shi,
  • Yuyang Xu,
  • Jianjun Xu

摘要

The Antarctic ozone hole constitutes a critical global environmental issue. Although signs of recovery have emerged since the early 21st century, ongoing climate change has heightened uncertainty about its future trajectory. This study uses ERA5 as a reference dataset and ozone simulations from 13 models participating in the Coupled Model Intercomparison Project Phase 6 (CMIP6), aiming to improve Antarctic ozone projections through the optimized model screening and Multi-Model Ensemble (MME) approach. Three novel SHAP (SHapley Additive exPlanations) value screening methods derived from Tree-Based models, along with one traditional screening method, are developed for CMIP6 ensemble selection. Based on these methods, five machine learning (ML) models are further constructed for future Antarctic ozone projections. The main conclusions are as follows: (1) All CMIP6 models show general consistency with ERA5 but exhibit varying degrees of bias. Even the best-performing model, CESM2, performs well in terms of Taylor Skill Score (TSS) and Root Mean Square Error (RMSE), but shows notable deficiencies in Interannual Variability Skill Score (IVS); (2) Among the four screening methods, the SHAP-based Worst-to-Best Screening, which progressively removes poorly performing models, achieves the best performance. The optimal Artificial Neural Network (ANN) model based on this method (MME6-ANN) significantly improves the MME (R = 0.833, IVS × 10 = 0.001, Mean Squared Error (MSE) = 200.81), substantially outperforming the traditional equal-weighted ensemble (MMM13) derived from 13 CMIP6 models and achieving the best agreement with the reference; (3) The Antarctic ozone recovery years projected by MME6-ANN under the four Shared Socioeconomic Pathway (SSP) scenarios (SSP1-2.6, SSP2-4.5, SSP3-7.0, and SSP5-8.5) are 2065, 2062, 2057, and 2056, respectively, representing delays of 9, 11, 7, and 7 years compared with the traditional MME. Overall, this study offers a more reliable estimate of future Antarctic ozone recovery under climate change, and provides a robust reference for CMIP6 MME projections using machine learning.

Graphical Abstract

The graphical abstract presents a machine learning–based enhanced CMIP6 model screening and multimoded ensemble framework (MME-ML) for improving Antarctic ozone future projections. ERA5 reanalysis data and ozone simulations from CMIP6 models were first processed into monthly time series, which are then divided into training, validation, and test sets. Subsequently, three innovative screening methods based on SHAP values derived from Tree-Based models, alongside one traditional screening method based on evaluation metrics including the TSS, IVS, and RMSE, were applied to screen the CMIP6 models. Based on the screened CMIP6 model ensembles from the four screening methods, five machine learning models (GLM, RF, XGBoost, LSTM, and ANN) were constructed, and a comprehensive evaluation identified the optimal screening method and its corresponding optimal ML model. Finally, the optimal ML model was used to improve the MME, correcting the CMIP6 Antarctic ozone projections under four SSP scenarios. The results indicate that the Antarctic ozone recovery years projected by the improved MME under the four scenarios (SSP1-2.6, SSP2-4.5, SSP3-7.0, and SSP5-8.5) are 2065, 2062, 2057, and 2056, respectively, representing delays of 9, 11, 7, and 7 years compared with the traditional MME. This research yields more accurate estimates of future Antarctic ozone recovery under climate change, and provides a transferable framework for CMIP6 multimodel ensemble projections using machine learning techniques.